Escolar Documentos
Profissional Documentos
Cultura Documentos
MODULE OF iNZight
Introduction
The Time Series module of the FREE software package iNZight uses two different statistical
models. The model used to obtain the series decomposition is called a Seasonal Trend
Lowess and the model used to calculate predictions is a Holt-Winters model. This guide is to
give teachers a brief summary of the models used but the new standard has no expectation
that students need to know any theoretical background to the models.
Smoothing or filtering a Time Series is best thought of as similar to the idea of filtering music
through an amplifier. We can amplify certain sounds or we can suppress certain sounds.
Similarly, we can suppress (remove) certain features in a Time Series, such as seasonality, in
order to model the trend and/or cycle. Once we have built a suitable model for the
smoothed series, we can add back the appropriate seasonal component in order to produce
predictions.
A common method for smoothing a Time Series is to use moving averages, which is what has
traditionally be taught in schools for AS 3.1. One drawback of moving averages is that our
moving average series becomes shorter than the original Time Series. If we have monthly
data, our first moving average value is calculated on observations 1 to 12, and the second
moving average value is calculated on observations 2 to 13. We then average these two
values to get our first moving average value which then replaces observation 7 in our original
series. Similarly, at the end of our series, there are six observations that we have no moving
average values for.
A more useful tool for isolating and then removing the seasonal component of a Time Series
is Seasonal Trend Lowess a decomposition function in R ( the programming language that
iNZight is written in). The method used is to first smooth the trend and cycle using a lowess
smoother (fitting a local regression to a window of points and using the point on the fitted
regression line as the value of the smooth for the time value in the middle of the window).
The regression that is used is weighted, in that observations near the edge of the window
are given less weight than observations near the centre of the window when determining
the local regression line. Then a separate lowess smoother is used on each seasonal sub-
series (i.e. all the January observations, all the February observations, ). The trend and
cycle smoothed value and the appropriate seasonal smoothed value can be subtracted
from the original observation to yield the remainder or random component for that
A third option for smoothing data is exponential smoothing and it is this technique that is
used in the Holt-Winters model.
Holt-Winters Model
This model, often referred to as a procedure, was first proposed in the early 1960s. It uses a
process known as exponential smoothing. All data values in a series contribute to the
calculation of the prediction model.
0
-2
-4
-6
Time
Exponential smoothing in its simplest form should only be used for non-seasonal time series
exhibiting a constant trend (or what is known as a stationary time series). It seems a
reasonable assumption to give more weight to the more recent data values and less weight
to the data values from further in the past. An intuitive set of weights is the set of weights
that decrease each time by a constant ratio. Strictly speaking this implies an infinite number
of past observations but in practice there will be a finite number. Such a procedure is known
as exponential smoothing since the weights lie on an exponential curve.
a denotes the smoothing parameter, the exponential smoothing constant, 0< <1
where S1 = y1
14 24 5 18 10 17 23 17 23
S1 = y1 = 14
etc
Thus the smoothed series depends on all previous values, with the most weight given to the
most recent values.
The Holt-Winters model uses a modified form of exponential smoothing. It applies three
exponential smoothing formulae to the series. Firstly, the level (or mean) is smoothed to
give a local average value for the series. Secondly, the trend is smoothed and lastly each
seasonal sub-series ( ie all the January values, all the February values.. for monthly data) is
smoothed separately to give a seasonal estimate for each of the seasons. A combination of
these three series is used to calculate the predictions output by iNZight.
The exponential smoothing formulae applied to a series with a trend and constant seasonal
component using the Holt-Winters additive technique are:
bt = ( at a t1 )+(1 )bt1
s t =(Y t at )+(1 )s t p
where:
The Holt-Winters algorithm requires starting (or initialising) values. Most commonly:
1
a p = (Y 1 +Y 2 ++Y p )
p
b p=
p[
1 Y p+1 Y 1 Y p+2Y 2
p
+
p
++
Y p+ p Y p
p ]
s 1 =Y 1a p , s2 =Y 2 a p , , s p =Y p a p
The Holt-Winters forecasts are then calculated using the latest estimates from the
appropriate exponential smooths that have been applied to the series.
^y T+ =aT + bT +s T
As mentioned earlier the Holt-Winters model assumes that the seasonal pattern is relatively
constant over the time period. Students would be expected to notice changes in the
seasonal pattern and identify this as a potential problem with the model, particularly if long
term predictions are made. In practice this is dealt with by transforming the original data
and modelling the transformed series or using a multiplicative model. Students are not
expected to know this, but are required to identify a variable seasonal pattern as a potential
problem. The exponential smoothing formulae applied to a series using Holt-Winters
Multiplicative models are:
Yt
at = +(1 )(at1 +b t1 )
s t p
bt = ( at a t1 )+(1 )bt1
Yt
s t = +(1 )s tp
at
Y1 Y2 Yp
s1= , s 2= , , sp=
ap ap ap
^y T+ =(aT + bT )sT
Holland.
There are many situations where it is important to give interval predictions, rather than
recommended that the term prediction interval is used in the context of time series analysis.
This is because prediction interval is more descriptive and because the term confidence
relatively common to see predictions made without any reference to prediction intervals.
This may be because there are a number of different ways that prediction intervals can be
calculated. The paper above provides not only details of how the prediction intervals for
Holt-Winters are produced but also compares the authors preferred method with several
alternative methods. It also compares the prediction intervals calculated for the same data
Derivation and details of prediction interval calculation can be found in Yar & Chatfields (1990) paper
see section 3 and 4 on page 129.
In one example given in the paper, a monthly index of employment in manufacturing in Canada, a
prediction for three years after the end of the actual data is provided of 115.9. A prediction interval
of [113.95,117.85] is also calculated. A suggested interpretation of this prediction interval (P.I.) is
There is a 95% chance that the true index value of employment in manufacturing in Canada in three
years time will be between 113.95 and 117.85.
The details given in this paper apply to an additive Holt-Winters model only.
The test of any prediction model is how well does it predict when compared to actual data
values. To do this either remove the last few given observations or find the next few actual
observations. Different prediction models can then be compared using a statistic known as
the Root Mean Squared Error of Prediction (RMSEP). The formula for calculating this statistic
is given below
Rachel Passmore, Royal Society Endeavour Teacher Fellow,
Department of Statistics, University of Auckland
1
RMSEP= ( y t y^ t )2
t=1
Want to know more? There are several Youtube clips that explain exponential smoothing
and Holt-Winters using Excel or R if you are interested.
Rachel Passmore
Email : passm@vodafone.co.nz
August 2012